4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 #include <linux/sched.h>
39 #include <linux/quotaops.h>
41 #define DEBUG_SUBSYSTEM S_LLITE
43 #include <obd_support.h>
44 #include <lustre_lite.h>
45 #include <lustre/lustre_idl.h>
46 #include <lustre_dlm.h>
48 #include "llite_internal.h"
50 static void free_dentry_data(struct rcu_head *head)
52 struct ll_dentry_data *lld;
54 lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
58 /* should NOT be called with the dcache lock, see fs/dcache.c */
59 static void ll_release(struct dentry *de)
61 struct ll_dentry_data *lld;
65 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
69 ll_intent_release(lld->lld_it);
70 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
72 LASSERT(lld->lld_cwd_count == 0);
73 LASSERT(lld->lld_mnt_count == 0);
75 call_rcu(&lld->lld_rcu_head, free_dentry_data);
80 /* Compare if two dentries are the same. Don't match if the existing dentry
81 * is marked invalid. Returns 1 if different, 0 if the same.
83 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
84 * an AST before calling d_revalidate_it(). The dentry still exists (marked
85 * INVALID) so d_lookup() matches it, but we have no lock on it (so
86 * lock_match() fails) and we spin around real_lookup(). */
87 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
88 const struct dentry *dentry, const struct inode *inode,
89 unsigned int len, const char *str, const struct qstr *name)
96 if (memcmp(str, name->name, len))
99 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
100 name->len, name->name, dentry, dentry->d_flags,
103 /* mountpoint is always valid */
104 if (d_mountpoint((struct dentry *)dentry))
107 if (d_lustre_invalid(dentry))
113 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
116 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
117 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
118 return LDLM_ITER_CONTINUE;
119 return LDLM_ITER_STOP;
122 /* find any ldlm lock of the inode in mdc and lov
126 static int find_cbdata(struct inode *inode)
128 struct ll_sb_info *sbi = ll_i2sbi(inode);
129 struct lov_stripe_md *lsm;
134 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
135 return_if_equal, NULL);
139 lsm = ccc_inode_lsm_get(inode);
143 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
144 ccc_inode_lsm_put(inode, lsm);
150 * Called when last reference to a dentry is dropped and dcache wants to know
151 * whether or not it should cache it:
152 * - return 1 to delete the dentry immediately
153 * - return 0 to cache the dentry
154 * Should NOT be called with the dcache lock, see fs/dcache.c
156 static int ll_ddelete(const struct dentry *de)
161 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
162 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
163 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
164 d_unhashed((struct dentry *)de) ? "" : "hashed,",
165 list_empty(&de->d_subdirs) ? "" : "subdirs");
167 /* kernel >= 2.6.38 last refcount is decreased after this function. */
168 LASSERT(d_count(de) == 1);
170 /* Disable this piece of code temproarily because this is called
171 * inside dcache_lock so it's not appropriate to do lots of work
172 * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
175 /* if not ldlm lock for this inode, set i_nlink to 0 so that
176 * this inode can be recycled later b=20433 */
177 if (de->d_inode && !find_cbdata(de->d_inode))
178 clear_nlink(de->d_inode);
181 if (d_lustre_invalid((struct dentry *)de))
186 static int ll_set_dd(struct dentry *de)
191 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
192 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
195 if (de->d_fsdata == NULL) {
196 struct ll_dentry_data *lld;
199 if (likely(lld != NULL)) {
200 spin_lock(&de->d_lock);
201 if (likely(de->d_fsdata == NULL))
205 spin_unlock(&de->d_lock);
214 int ll_dops_init(struct dentry *de, int block, int init_sa)
216 struct ll_dentry_data *lld = ll_d2d(de);
219 if (lld == NULL && block != 0) {
227 if (lld != NULL && init_sa != 0)
228 lld->lld_sa_generation = 0;
230 /* kernel >= 2.6.38 d_op is set in d_alloc() */
231 LASSERT(de->d_op == &ll_d_ops);
235 void ll_intent_drop_lock(struct lookup_intent *it)
237 if (it->it_op && it->d.lustre.it_lock_mode) {
238 struct lustre_handle handle;
240 handle.cookie = it->d.lustre.it_lock_handle;
242 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
243 " from it %p\n", handle.cookie, it);
244 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
246 /* bug 494: intent_release may be called multiple times, from
247 * this thread and we don't want to double-decref this lock */
248 it->d.lustre.it_lock_mode = 0;
249 if (it->d.lustre.it_remote_lock_mode != 0) {
250 handle.cookie = it->d.lustre.it_remote_lock_handle;
252 CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
253 LPX64" from it %p\n", handle.cookie, it);
254 ldlm_lock_decref(&handle,
255 it->d.lustre.it_remote_lock_mode);
256 it->d.lustre.it_remote_lock_mode = 0;
261 void ll_intent_release(struct lookup_intent *it)
265 CDEBUG(D_INFO, "intent %p released\n", it);
266 ll_intent_drop_lock(it);
267 /* We are still holding extra reference on a request, need to free it */
268 if (it_disposition(it, DISP_ENQ_OPEN_REF))
269 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
270 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
271 ptlrpc_req_finished(it->d.lustre.it_data);
272 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
274 ptlrpc_req_finished(it->d.lustre.it_data);
276 it->d.lustre.it_disposition = 0;
277 it->d.lustre.it_data = NULL;
281 void ll_invalidate_aliases(struct inode *inode)
283 struct dentry *dentry;
284 struct ll_d_hlist_node *p;
287 LASSERT(inode != NULL);
289 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
290 inode->i_ino, inode->i_generation, inode);
292 ll_lock_dcache(inode);
293 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
294 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
295 "inode %p flags %d\n", dentry->d_name.len,
296 dentry->d_name.name, dentry, dentry->d_parent,
297 dentry->d_inode, dentry->d_flags);
299 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
300 CERROR("called on root (?) dentry=%p, inode=%p "
301 "ino=%lu\n", dentry, inode, inode->i_ino);
302 lustre_dump_dentry(dentry, 1);
303 libcfs_debug_dumpstack(NULL);
306 d_lustre_invalidate(dentry, 0);
308 ll_unlock_dcache(inode);
313 int ll_revalidate_it_finish(struct ptlrpc_request *request,
314 struct lookup_intent *it,
323 if (it_disposition(it, DISP_LOOKUP_NEG))
326 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
331 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
334 LASSERT(dentry != NULL);
336 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
337 struct inode *inode = dentry->d_inode;
338 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
340 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
341 inode, inode->i_ino, inode->i_generation);
342 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
345 /* drop lookup or getattr locks immediately */
346 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
347 /* on 2.6 there are situation when several lookups and
348 * revalidations may be requested during single operation.
349 * therefore, we don't release intent here -bzzz */
350 ll_intent_drop_lock(it);
354 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
356 struct lookup_intent *it = *itp;
358 if (!it || it->it_op == IT_GETXATTR)
363 int ll_revalidate_it(struct dentry *de, int lookup_flags,
364 struct lookup_intent *it)
366 struct md_op_data *op_data;
367 struct ptlrpc_request *req = NULL;
368 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
369 struct obd_export *exp;
370 struct inode *parent = de->d_parent->d_inode;
374 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
377 if (de->d_inode == NULL) {
380 /* We can only use negative dentries if this is stat or lookup,
381 for opens and stuff we do need to query server. */
382 /* If there is IT_CREAT in intent op set, then we must throw
383 away this negative dentry and actually do the request to
384 kernel to create whatever needs to be created (if possible)*/
385 if (it && (it->it_op & IT_CREAT))
388 if (d_lustre_invalid(de))
391 ibits = MDS_INODELOCK_UPDATE;
392 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
396 /* Never execute intents for mount points.
397 * Attributes will be fixed up in ll_inode_revalidate_it */
398 if (d_mountpoint(de))
399 GOTO(out_sa, rc = 1);
401 /* need to get attributes in case root got changed from other client */
402 if (de == de->d_sb->s_root) {
403 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
409 exp = ll_i2mdexp(de->d_inode);
411 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
412 ll_frob_intent(&it, &lookup_it);
415 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
418 if (it->it_op == IT_OPEN) {
419 struct inode *inode = de->d_inode;
420 struct ll_inode_info *lli = ll_i2info(inode);
421 struct obd_client_handle **och_p;
426 * We used to check for MDS_INODELOCK_OPEN here, but in fact
427 * just having LOOKUP lock is enough to justify inode is the
428 * same. And if inode is the same and we have suitable
429 * openhandle, then there is no point in doing another OPEN RPC
430 * just to throw away newly received openhandle. There are no
431 * security implications too, if file owner or access mode is
432 * change, LOOKUP lock is revoked.
436 if (it->it_flags & FMODE_WRITE) {
437 och_p = &lli->lli_mds_write_och;
438 och_usecount = &lli->lli_open_fd_write_count;
439 } else if (it->it_flags & FMODE_EXEC) {
440 och_p = &lli->lli_mds_exec_och;
441 och_usecount = &lli->lli_open_fd_exec_count;
443 och_p = &lli->lli_mds_read_och;
444 och_usecount = &lli->lli_open_fd_read_count;
446 /* Check for the proper lock. */
447 ibits = MDS_INODELOCK_LOOKUP;
448 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
450 mutex_lock(&lli->lli_och_mutex);
451 if (*och_p) { /* Everything is open already, do nothing */
452 /*(*och_usecount)++; Do not let them steal our open
453 handle from under us */
454 SET_BUT_UNUSED(och_usecount);
455 /* XXX The code above was my original idea, but in case
456 we have the handle, but we cannot use it due to later
457 checks (e.g. O_CREAT|O_EXCL flags set), nobody
458 would decrement counter increased here. So we just
459 hope the lock won't be invalidated in between. But
460 if it would be, we'll reopen the open request to
461 MDS later during file open path */
462 mutex_unlock(&lli->lli_och_mutex);
465 mutex_unlock(&lli->lli_och_mutex);
469 if (it->it_op == IT_GETATTR) {
470 rc = ll_statahead_enter(parent, &de, 0);
473 else if (rc != -EAGAIN && rc != 0)
478 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
479 de->d_name.name, de->d_name.len,
480 0, LUSTRE_OPC_ANY, NULL);
482 RETURN(PTR_ERR(op_data));
484 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
485 it->it_create_mode &= ~current_umask();
486 it->it_create_mode |= M_CHECK_STALE;
487 rc = md_intent_lock(exp, op_data, NULL, 0, it,
489 &req, ll_md_blocking_ast, 0);
490 it->it_create_mode &= ~M_CHECK_STALE;
491 ll_finish_md_op_data(op_data);
493 /* If req is NULL, then md_intent_lock only tried to do a lock match;
494 * if all was well, it will return 1 if it found locks, 0 otherwise. */
495 if (req == NULL && rc >= 0) {
503 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
504 "%d\n", rc, it->d.lustre.it_status);
510 rc = ll_revalidate_it_finish(req, it, de);
512 if (rc != -ESTALE && rc != -ENOENT)
513 ll_intent_release(it);
517 if ((it->it_op & IT_OPEN) && de->d_inode &&
518 !S_ISREG(de->d_inode->i_mode) &&
519 !S_ISDIR(de->d_inode->i_mode)) {
520 ll_release_openhandle(de, it);
525 /* We do not free request as it may be reused during following lookup
526 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
527 * be freed in ll_lookup_it or in ll_intent_release. But if
528 * request was not completed, we need to free it. (bug 5154, 9903) */
529 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
530 ptlrpc_req_finished(req);
532 /* mdt may grant layout lock for the newly created file, so
533 * release the lock to avoid leaking */
534 ll_intent_drop_lock(it);
535 ll_invalidate_aliases(de->d_inode);
538 __u64 matched_bits = 0;
540 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
541 "inode %p refc %d\n", de->d_name.len,
542 de->d_name.name, de, de->d_parent, de->d_inode,
545 ll_set_lock_data(exp, de->d_inode, it, &bits);
547 /* Note: We have to match both LOOKUP and PERM lock
548 * here to make sure the dentry is valid and no one
549 * changing the permission.
550 * But if the client connects < 2.4 server, which will
551 * only grant LOOKUP lock, so we can only Match LOOKUP
552 * lock for old server */
553 if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
554 OBD_CONNECT_LVB_TYPE)
556 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
558 matched_bits = MDS_INODELOCK_LOOKUP;
560 if (((bits & matched_bits) == matched_bits) &&
561 d_lustre_invalid(de))
562 d_lustre_revalidate(de);
563 ll_lookup_finish_locks(it, de);
567 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
568 ll_statahead_mark(parent, de);
572 * This part is here to combat evil-evil race in real_lookup on 2.6
573 * kernels. The race details are: We enter do_lookup() looking for some
574 * name, there is nothing in dcache for this name yet and d_lookup()
575 * returns NULL. We proceed to real_lookup(), and while we do this,
576 * another process does open on the same file we looking up (most simple
577 * reproducer), open succeeds and the dentry is added. Now back to
578 * us. In real_lookup() we do d_lookup() again and suddenly find the
579 * dentry, so we call d_revalidate on it, but there is no lock, so
580 * without this code we would return 0, but unpatched real_lookup just
581 * returns -ENOENT in such a case instead of retrying the lookup. Once
582 * this is dealt with in real_lookup(), all of this ugly mess can go and
583 * we can just check locks in ->d_revalidate without doing any RPCs
587 if (it != &lookup_it) {
588 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
589 if (it->it_op == IT_GETATTR)
590 lookup_it.it_op = IT_GETATTR;
591 ll_lookup_finish_locks(it, de);
595 /* Do real lookup here. */
596 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
597 de->d_name.len, 0, (it->it_op & IT_CREAT ?
599 LUSTRE_OPC_ANY), NULL);
601 RETURN(PTR_ERR(op_data));
603 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
604 ll_md_blocking_ast, 0);
606 struct mdt_body *mdt_body;
607 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
608 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
611 fid = *ll_inode2fid(de->d_inode);
613 /* see if we got same inode, if not - return error */
614 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
615 ll_finish_md_op_data(op_data);
617 goto revalidate_finish;
619 ll_intent_release(it);
621 ll_finish_md_op_data(op_data);
626 * For rc == 1 case, should not return directly to prevent losing
627 * statahead windows; for rc == 0 case, the "lookup" will be done later.
629 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
630 ll_statahead_enter(parent, &de, 1);
635 * Always trust cached dentries. Update statahead window if necessary.
637 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
639 struct inode *parent = dentry->d_parent->d_inode;
643 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
644 dentry->d_name.name, flags);
646 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
647 ll_need_statahead(parent, dentry) > 0) {
648 if (flags & LOOKUP_RCU)
651 if (dentry->d_inode == NULL)
653 do_statahead_enter(parent, &dentry, unplug);
654 ll_statahead_mark(parent, dentry);
661 void ll_d_iput(struct dentry *de, struct inode *inode)
664 if (!find_cbdata(inode))
669 struct dentry_operations ll_d_ops = {
670 .d_revalidate = ll_revalidate_nd,
671 .d_release = ll_release,
672 .d_delete = ll_ddelete,
674 .d_compare = ll_dcompare,